Acquisition-driven domain shift

An MRI domain is a bundle of choices

Suppose an image model is trained on knee MRI acquired at one field strength, with one vendor’s hardware and one protocol, and is then evaluated on images made under another combination. A change in the displayed anatomy may reflect the patient, but a change in intensity, noise, distortion, or texture can also come from image formation itself. The model is not seeing a new patient on the same measurement system; it is seeing a different input distribution.

In MRI, this problem is visible because ordinary intensity is relative and many acquisition decisions affect the signal at once. The relevant domain is therefore a bundle of choices rather than a single scanner label. The axes below are useful questions to record and, where the claim requires it, to vary deliberately in evaluation.

Axis	What can change
Field strength	relaxation behaviour, susceptibility effects, signal and contrast conditions
Scanner and vendor	magnet, gradient, receiver, calibration, and implementation characteristics
Receive coil	sensitivity profile, coverage, and the spatial pattern of signal and noise
Sequence and timing	TR, TE, flip angle, TI, echo train, and other preparation/readout choices
Protocol	resolution, field of view, orientation, acceleration, fat suppression, and b-values
Sampling and reconstruction	trajectory, parallel imaging, compressed sensing, learned reconstruction, and regularisation
Site and workflow	local calibration, positioning, operator practice, and other environment-specific conditions

These axes interact. Changing a protocol may change both sampling and contrast; changing the coil changes the signal field that the reconstruction receives; changing field strength can alter the relationship between tissue relaxation and the chosen timing. Consequently, “same sequence name” is not enough to establish an identical measurement condition. Patient mix, disease prevalence, and clinical indication can also shift a dataset, but those population shifts should be described separately from the acquisition-driven changes listed here.

Domain shift changes the scope of a claim

The practical error is not that a model has seen one protocol. The error is to treat performance within that protocol as evidence for every other protocol. A model trained on one combination of field strength, vendor, coil, sequence, and reconstruction may learn anatomy, but it may also use stable acquisition cues that do not survive at the deployment site. This can affect segmentation, detection, classification, radiomics, and synthesis in different ways.

The remedy begins with documentation. A study should describe the acquisition axes in both development and evaluation data, including the axes that were held fixed. If the claim is transport across vendors or field strengths, the evaluation should contain a meaningful change on that axis rather than only a random split from the original domain. When a deployment claim is being tested, a frozen model and a prespecified operating rule make the effect of the changed input distribution visible; any later adaptation should be reported as a separate procedure rather than folded into the original test.

Normalisation and harmonisation can be useful processing choices, but they do not create a universal physical intensity scale or guarantee that all acquisition differences have been removed. Their parameters and fitting data belong in the provenance, and their application must not leak information from an evaluation set into model development. The right question is not whether a dataset has been made to look uniform, but which differences remain and whether the intended claim was tested across them.

This is the MRI-specific version of the broader validation discipline: a result is only as general as the acquisition and population changes it has survived. The next section applies the same boundary to image synthesis, where a model is asked to produce a target contrast that may not have been measured at all.

Sources

• Abi Berger, “Magnetic resonance imaging” — the dependence of MR signal and contrast on magnetic/RF preparation and pulse-sequence conditions.
• Wiltgen T, Voon C, Van Leemput K, Wiestler B, Muehlau M. Intensity scaling of conventional brain magnetic resonance images avoiding cerebral reference regions: A systematic review. PLoS One. 2024;19:e0298642. doi:10.1371/journal.pone.0298642. (Full text)
• Kushol R, Wilman AH, Kalra S, Yang YH. DSMRI: Domain Shift Analyzer for Multi-Center MRI Datasets. Diagnostics. 2023;13:2947. doi:10.3390/diagnostics13182947. (Full text)
• Riley RD, Archer L, Snell KIE, et al. Evaluation of clinical prediction models (part 2): how to undertake an external validation study. BMJ. 2024;384:e074820. doi:10.1136/bmj-2023-074820. (Full text)